The next big threat to Nvidia Corp. (NVDA) dominance may not come from another GPU maker. It could come from the companies buying Nvidia's technology and deciding they want more control over what runs inside their AI data centers — a shift that has put Marvell Technology, Inc. (MRVL) on a path toward a potentially $30 billion custom-chip business.

Hyperscalers Want More Control

The AI infrastructure race is increasingly moving beyond the question of who makes the fastest accelerator. The biggest cloud companies are also trying to determine which parts of their AI systems should be built specifically for their own workloads, particularly as inference — running trained AI models — becomes a larger share of computing demand.

That is where Marvell sees an opening.

At its Investor Day, Marvell said it is already shipping custom silicon to all four major U.S. hyperscalers. Its custom business generated about $1.5 billion last year, but the company now expects revenue from the business to exceed $12 billion in fiscal 2029, up from its previous target of more than $10 billion.

CEO Matt Murphy described the opportunity as a progression: hyperscalers can start with Marvell's standard products, move toward customized systems and eventually commission fully custom silicon incorporating their own intellectual property.

That matters because Marvell is not asking customers to abandon their existing AI infrastructure. It is becoming the company that helps them tailor it.

The $30 Billion Opportunity

Marvell now expects its custom business to reach approximately $30 billion in fiscal 2031, with the opportunity becoming more balanced between its XPU products and what it calls "XPU attach" — networking, memory, storage and other components that connect to AI accelerators.

Management says the forecast is based largely on programs and customer positions already in place rather than a need for a wave of entirely new wins. Marvell's custom chief Will Chu said the company has momentum across all tier-one hyperscalers and expects higher attach rates as AI systems become more complex.

The Alphabet Inc. (GOOGL) (GOOG)/Google relationship shows why this market is becoming strategically important. In August, Marvell disclosed a broad agreement covering custom products tied to Google's TPU ecosystem, including inference accelerators, storage and networking controllers and memory interfaces.

The agreement could generate up to roughly $120 billion in revenue through fiscal 2033 if performance targets are met.

Nvidia Isn't Necessarily Losing

There is an important twist to the Nvidia angle: Marvell's rise does not automatically mean Nvidia loses.

In fact, the companies announced a strategic partnership in March under Nvidia's NVLink Fusion platform. Marvell can provide custom XPUs and networking compatible with Nvidia's architecture, allowing customers to customize parts of their systems while retaining Nvidia technology elsewhere. Nvidia also invested $2 billion in Marvell.

That makes Marvell's opportunity more interesting than a simple Nvidia-versus-Marvell contest. The company can benefit from hyperscalers wanting proprietary silicon while remaining connected to the broader Nvidia ecosystem.

For Marvell investors, the key question is no longer whether AI spending will grow. It is whether the shift toward customized AI infrastructure can turn today's design wins into the company's projected $30 billion custom business.

For Nvidia investors, the more important signal may be how much of future AI spending remains concentrated in GPUs — and how much migrates toward the custom silicon and connectivity layer around them.